Application of Machine Learning to Identify and Characterise Thermochemical Heterogeneity in the Earth's Lower Mantle
Abstract
3-D variations of wave speeds and density have shown the presence of seismically distinct structures in the Earth's mantle. However, seismology does not directly provide information on the thermochemical properties of these structures, which is crucial for understanding mantle dynamics and thus, planetary evolution and origin. Comparing observed seismic data with mineral properties obtained from mineral physics allows us to make thermochemical interpretations of seismic observations. Inverting seismic properties for thermochemical properties is challenging because of the non-uniqueness of the problem: different combinations of temperature and composition can produce the same seismic properties.
To tackle this, we solve the inversion as a probabilistic inverse problem using neural networks. We present one of the first applications of machine learning for interpreting seismic observables in terms of thermochemical parameters. This approach allows us to efficiently analyse the uncertainties and underlying trade-offs among the various parameters. By doing so, we explore the following: a) what can we say about the temperature and bulk composition of the mantle using seismic wave speeds (+/- density), and b) to what extent are the trade-offs inferred using wave speeds reduced by including density? Neural networks are trained with wave speeds (+/- density) as inputs which deliver probability density functions of temperature, mineralogy and bulk composition as outputs. Applying the trained neural networks to a recent long-wavelength tomographic model identifies and characterises thermochemical heterogeneity in the lower mantle. Using wave speeds alone, we can constrain variations in the SiO2 content and also distinguish seismically fast tectonic slabs from slow "piles" in terms of temperature but with large uncertainty. The combination of wave speeds and density further constrains FeO and MgO contents as well as reduces the uncertainty on the temperature and Si, by breaking down trade-offs between thermochemical parameters. When there are variations in Fe content, we may infer the wrong temperature if we do not have a constraint on the density. The dense slow piles at the base of the lower mantle can be explained by enrichment in Fe and Si- a characteristic feature of enstatite chondritic compositions.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMDI25A..08R